ICCV 2025 — visual examples

64 of 64 oral records have a writing-focused manuscript study. Official program snapshot: 2026-10-07. Selected rendered pages were inspected for each study; this is not a proof audit or reproduction of experiments.

The latest completed ICCV oral corpus is the 2025 edition: 64 distinct Oral events across 12 sessions, matched to CVF proceedings. The feed contains 128 oral-decision events because each paper also has a poster record; those are not 128 papers.

The writing corpus uses the 64 official main PDFs. Separate supporting links exist for 59 papers; five have no linked supplement. Those separate assets are not claimed as inspected appendices, and unavailable measurements are not replaced with zeros.

Writing-focused manuscript scans and selected raster-page checks are not line-by-line proof audits or experiment reproduction. Published metadata can differ from event metadata; both are retained where relevant. ICCV 2027 is announced but is not a completed studied edition.

Inspect the complete inventory and source status · Official oral program

Choose a visual for its explanatory job

Use a diagram to explain a mechanism or assumption, a plot to expose a relationship, a table for exact comparison, and aligned qualitative panels for visible behavior. Captions should define the setting, symbols, units and supported conclusion. Match the display to the evidence, not to decoration.

Read the display, then its design lesson. Tables and numeric charts below reconstruct explicitly attributed factual subsets. Diagrams are original teaching layouts, not screenshots or reproductions of the papers' architecture. Selected examples are not a prevalence ranking of visual styles.

Data budget is a separate intervention

GT-Loc: Unifying When and Where in Images Through a Joint Embedding Space · Table 4, physical PDF p.7

Training data (%)	Month error (lower)	Hour error (lower)	TPS (higher)
100	1.4	2.72	77
50	1.69	2.83	74.02
10	1.7	2.94	73.51
5	1.89	2.86	72.07

Source-reported numerical subset. GT-Loc time prediction, unseen SkyFinder cameras; fraction of training data varied. Source reports no uncertainty for these rows.

Observed in the source: A data-budget table makes the degradation visible; retain the non-monotonic hour-error row.

Prepare: State the evaluation setting, units, metric directions and subset selection directly beside the display.

Do not infer: Do not repeat the adjacent prose's understated month-error increase; the displayed 100%-to-5% change is 0.49 months, not 0.3.

An aggregate leader need not win every task

Scaling Laws for Native Multimodal Models · Table 4, physical PDF p.7 (printed p.18)

Model	TextVQA accuracy	COCO CIDEr
Late-fusion	25.8	70.7
Early-fusion	28.1	72
Early-MoEs	30	69.6

Source-reported numerical subset. Native models at 1.5B scale, pretrained on 300B tokens and supervised-finetuned on the LLaVA mixture. Only two columns of the source table are reproduced.

Observed in the source: Group incompatible metrics and retain trade-offs instead of coloring an entire winning row.

Prepare: State the evaluation setting, units, metric directions and subset selection directly beside the display.

Do not infer: This is a selected two-task subset, not the paper's aggregate score or a broad ranking; the source gives no uncertainty for these cells.

Name the non-winning endpoint

Differentiable Room Acoustic Rendering with Multi-View Vision Priors · Table 2, physical PDF p.7 (printed p.43)

Method	Loudness error (dB, lower)	C50 error (dB, lower)	T60 error (%, lower)
AV-NeRF	1.26	1.03	9.5
Diff-RIR	1.32	3.13	6.8
AV-DAR	0.85	1.15	6.3

Source-reported numerical subset. Hearing Anything Anywhere, Hallway room only; 2.0s RIR segments at 16kHz; trained on 12 listener locations. Three selected methods and all three displayed Hallway metrics.

Observed in the source: The results prose explicitly acknowledges the AV-NeRF C50 advantage; a selective reconstruction should preserve it.

Prepare: State the evaluation setting, units, metric directions and subset selection directly beside the display.

Do not infer: Single-room subset; not an average across all six real environments and not evidence of independent material-property validation.

Pair quality with resource cost

Removing Cost Volumes from Optical Flow Estimators · Table 5, physical PDF p.7 (printed p.85); protocol physical p.6

Axes focus on the displayed range, not a zero baseline. Point numbers match the named configurations below the plot; exact values and comparison conditions remain in the table and caption.

Method	Total EPE (lower)	Runtime (ms, lower)	Memory (GB, lower)
SEA-RAFT	0.54	169.72	8.22
ReCoVEr-MN	0.99	50.16	0.49
ReCoVEr-CX	0.51	144.41	1.24

Source-reported numerical subset. Spring training split used only for evaluation, full-resolution frames without resizing; all three selected rows mark Sintel and TartanAir training. Runtime uses batch1, FP32, 10 warmup and 100 timed iterations. The main text names training GPUs but does not separately unambiguously identify runtime hardware, so do not invent it.

Observed in the source: The small model exchanges accuracy for much lower memory/time; the larger model occupies a different operating point.

Prepare: State the evaluation setting, units, metric directions and subset selection directly beside the display.

Do not infer: Three-row subset; no error bars supplied. Do not equate training-data flags with identical architecture or tuning budgets.

Make a cyclic variable operational

GT-Loc: Unifying When and Where in Images Through a Joint Embedding Space · Figures 2–3, physical p.4; Section3.2, p.5

Original teaching construction. An original 23:00-to-01:00 example with a 24-hour period. It illustrates wraparound, not a result or the source artwork; the complete method also treats the month coordinate.

Observed in the source: A tiny numerical example can explain a representation choice before the full objective.

Prepare: Name the coordinate period and compare naive versus wrapped distance before the full objective.

Do not infer: Do not copy the published grid, colors, geometry or any result values; this is an explanatory construction, not a reproduced experiment.

Explain overlap through ordered boundary events

EVER: Exact Volumetric Ellipsoid Rendering for Real-time View Synthesis · Figure3 and Equation2, physical p.3

Original teaching construction. An original simplified equal-density interval example motivated by the source boundary-event explanation. Color weighting, transmittance and full rendering are omitted.

Observed in the source: For a stateful exact procedure, show what changes at each event rather than only drawing a high-level renderer box.

Prepare: Label entry/exit events and the state changed at each boundary before presenting the full procedure.

Do not infer: This teaching construction omits color weighting and transmittance; it must not imply the staircase alone is the full rendering algorithm.